GpuCoopMat

1 min read

Import: import system.gpu;

This reference lists the public declarations for this part of the library. See Reading API signatures for parameter and result notation. Shader-only declarations can be used only inside @compute, @vertex, @fragment, inferred helpers and ray tracing shader functions; using them elsewhere is a compile error.

GpuCoopMat

A matrix tile that a subgroup of threads holds and multiplies together on tensor hardware, for compute shaders that tile their own matrix multiplications. Its size is fixed when the shader is built from the constant sizes passed to zeros or load. Needs GpuFeature.CooperativeMatrices, which GpuTarget.Vulkan, Metal and Cuda can provide. Metal tiles are 8 by 8, so a shader that targets GpuTarget.Metal must use that size. Using it with only other targets is a build error.

public struct GpuCoopMat<T>

zeros

A rows by columns tile of zeros; both sizes must be constants.

public static extern fun zeros(rows: int, columns: int): GpuCoopMat<T>

load

Loads a rows by columns tile whose first element is buffer[offset] and whose rows are row_stride elements apart. The sizes must be constants.

public static extern fun load(
    buffer: GpuBuffer<T>,
    offset: int,
    row_stride: int,
    rows: int,
    columns: int
): GpuCoopMat<T>

multiply_add

a * b + c, where a is M by K, b K by N, and c and the result M by N.

public static extern fun multiply_add(
    a: GpuCoopMat<T>,
    b: GpuCoopMat<T>,
    c: GpuCoopMat<T>
): GpuCoopMat<T>

store

Stores the tile into buffer starting at buffer[offset], rows row_stride elements apart.

public extern fun store(buffer: GpuBuffer<T>, offset: int, row_stride: int): void